53 comments

[ 1.8 ms ] story [ 79.4 ms ] thread
LLMs on device is the future. It's more secure and solves the problem of too much demand for inference compared to data center supply, it also would use less electricity. It's just a matter of getting the performance good enough. Most users don't need frontier model performance.
still waiting for the day I can comfortably run Claude Code with local llm's on MacOS with only 16gb of ram
Already running qwen 70b 4-bit on m2 max 96gb through llama.cpp and it's pretty solid for day to day stuff. The mlx switch is interesting because ollama was basically shelling out to llama.cpp on mac before, so native mlx should mean better memory handling on apple silicon. Curious to see how it compares on the bigger models vs the gguf path
The key difference is that MLX's array model assumes unified memory from the ground up. llama.cpp's Metal backend works fine but carries abstractions from the discrete GPU world — explicit buffer synchronization, command buffer boundaries — that are unnecessary when CPU and GPU share the same address space. You'll notice the gap most at large context lengths where KV cache pressure is highest.
"We can run your dumbed down models faster":

#The use of NVFP4 results in a 3.5x reduction in model memory footprint relative to FP16 and a 1.8x reduction compared to FP8, while maintaining model accuracy with less than 1% degradation on key language modeling tasks for some models.

What is the difference between Ollama, llama.cpp, ggml and gguf?
How does this compare to llama.cpp in terms of performance?
(comment deleted)
Finally! My local infra is waiting for it for months!
Good to see Ollama is catching up with the times for inference on Mac. MLX powered inference makes a big difference, especially on M5 as their graphs point out. What really has been a game changer for my workflow is using https://omlx.ai/ that has SSD KV cold caching. No longer have to worry about a session falling out of memory and needing to prefill again. Combine that with the M5 Max prefill speed means more time is spend on generation than waiting for 50k+ content window to process.
Yeah omlx seems to me like the front runner right now for running MLX models locally in agent workflows (which depend heavily on caching).
Why are people still using Ollama? Serious.

Lemonade or even llama.cpp are much better optimised and arguably just as easy to use.

I agree. In my experience on a macbook as well it can be terrifically hard to get some models to run properly on gpu w/ ollama, let alone containerised ollama.
I created "apfel" https://github.com/Arthur-Ficial/apfel a CLI for the apple on-device local foundation model (Apple intelligence) yeah its super limited with its 4k context window and super common false positives guardrails (just ask it to describe a color) ... bit still ... using it in bash scripts that just work without calling home / out or incurring extra costs feels super powerful.
This is quite interesting. I wonder if AFM is smart enough to do spam classification.
this is great! Incredibly fast and is working pretty well running loads on my m4 max studio.

Weirdly though I'm getting things like this: Apple FM is fast and free but has a hard limitation — it can't process prompts with Spanish/non-English words, which is a dealbreaker for California and Southwest real estate where half the street names are Spanish.

What would be the non Mac computer to run these models locally at the same performance profile? Any similar linux ARM based computers that can reach the same level?
> Please make sure you have a Mac with more than 32GB of unified memory.

Yeah, I can still save money by buying a cheaper device with less RAM and just paying my PPQ.AI or OpenRouter.com fees .

What are significant differences between Ollama and LM Studio now? I haven’t used Ollama because it was missing MLX when I started using LLM GUIs.
I have an M4 Max with 48GB RAM. Anyone have any tips for good local models? Context length? Using the model recommended in the blog post (qwen3.5:35b-a3b-coding-nvfp4) with Ollama 0.19.0 and it can take anywhere between 6-25 seconds for a response (after lots of thinking) from me asking "Hello world". Is this the best that's currently achievable with my hardware or is there something that can be configured to get better results?
I made my M2 Max generate a biryani recipe for me last night with 64gb ram and the baseline qwen3.5:35b model. I used the newest ollama with MLX.

https://gist.github.com/kylehotchkiss/8f28e6c75f22a56e8d2d31...

Under 3 minutes to get all that. The thinking is amusing, my laptop got quite warm, but for a 35b model on nearly 4 year old hardware, I see the light. This is the future.

The 35b-a3b-coding-nvfp4 model has the recommended hyperparameters set for coding, not chatting. If you want to use it to chat you can pull the `35b-a3b-nvfp4` model (it doesn't need to re-download the weights again so it will pull quickly) which has the presence penalty turned on which will stop it from thinking so much. You can also try `/set nothink` in the CLI which will turn off thinking entirely.
Get turboquant 4 bit implemented and this would be game changer.
> Please make sure you have a Mac with more than 32GB of unified memory. Time for an upgrade I guess. If I can run Qwen3.5 locally than it is time to switch over to local first LLM usage.
As being on the market for a new mac and comparing refub M4 Max vs M5 _Pro_, I am interested in how much faster the neural engines are -- compared to marketing claims.